The best contemporary machine-learning methods are exceedingly complex black boxes for performing classification tasks, given reliable, representative training data. Almost no astrophysics problems fall into this category! First, most astronomy problems are structured more like regressions than classifications (even discovery tasks). Second, our training sets tend to be labeled using other noisy data, or unreliable theory, not any kind of true values, so our supervised methods tend to look more like self-supervised methods or else emulators. Third, our training data sets are noisy, and they are never representative; we always want to find fainter, more distant, cooler, or weirder versions of the things we know. Fourth, our goals are not (usually) prediction in the space of the data; our goals are understanding, and understanding in the space of physical laws, theoretical models, or scientific principles. For all these reasons, we should be using machine learning in ways that are not all that similar to how it is used to organize your Twitter feed: We should be using (at first, anyway) methods that are simpler or more understandable and interpretable, such as linear models and Gaussian processes, which are easy to incorporate into probabilistic models. We should be using machine learning to address nuisances (like noise, calibration, foregrounds) or to speed up onerous computation. And we should hack machine learning methods to re-purpose their best internal sub-parts in the service of making better physics-inspired models that lead to new discoveries and new understandings. I give specific examples of successful work along all these lines.